Two years of noise have made this harder to answer than it should be. Here is the version we would give a client who asked over coffee.
Where it genuinely works today
Searching your own documents
If your business has years of contracts, manuals or notes, letting someone ask a question in plain language and get the answer with the source is real value. Retrieval keeps it grounded in your documents rather than the model’s imagination.
First drafts, never final ones
Useful for the blank page. Not useful for anything that goes out unread — the failure mode is confident, fluent and wrong, which is the hardest kind to catch.
Pulling structure out of mess
Invoices, emails, forms, PDFs. Turning unstructured input into structured data is unglamorous and genuinely saves hours, because it is the work nobody wanted.
Triage
Routing, tagging and prioritising incoming work. It does not have to be right every time to be useful — it has to be better than an unsorted queue.
Where it will embarrass you
- Anything requiring correctness without review. Prices, legal terms, medical or financial specifics. The model does not know when it is wrong.
- Customer-facing chat with no guardrails. If it can promise a refund policy you do not have, it eventually will.
- Decisions you cannot explain. If you cannot say why, you cannot defend it to a customer or a regulator.
How to start without wasting money
- Pick one task people currently do by hand, repeatedly, that takes real hours.
- Ship the smallest version to a handful of internal users.
- Measure whether it saved time — honestly, including the checking.
- Only then decide whether it deserves to be a product.
The projects that fail are the ones that begin with "we should use AI" instead of "this specific task is painful". The technology is fine. The framing is what sinks it.
